What Actually Shipped (Verified November 8, 2026)
Moonshot AI launched Kimi K3.1 Pro on October 23, 2026 and published the complete open weights on November 2: two days ahead of its own November 4 deadline. The Hugging Face repository (moonshotai/Kimi-K3.1-Pro) confirms the headline specs: a Mixture-of-Experts model with 2.8 trillion total parameters and 104 billion active per token, 896 experts with 16 selected per token, a 1,048,576-token context window, and vision input via a 401M-parameter MoonViT-V2 encoder. The download is about 1.56 TB across 96 shards.
The Kimi K3.1-Pro License lets you use, modify and sell the model, with two strings attached: hosted services earning over $20 million in revenue over 12 months need a separate agreement with Moonshot, and products exceeding 100 million monthly users or $20 million in monthly revenue must display the model's name. For everyone else (startups, research labs, and in-house deployments), it is effectively free frontier intelligence.
Get it (open weights): huggingface.co/moonshotai/Kimi-K3.1-Pro
Try it (web app): kimi.com
Pricing: API vs Self-Hosting
The API price cut is the practical story for most developers: K3.1 Pro costs a third less than K3 while scoring higher. Here is the full rate card (official pricing, checked November 8, 2026):
| Model / Channel | Input (per 1M) | Output (per 1M) | Cached input |
|---|---|---|---|
| kimi-k3.1-pro (API) | $2.00 | $10.00 | $0.20 |
| kimi-k3.1-pro-highspeed (API) | $4.00 | $20.00 | $0.20 |
| Kimi K3 (legacy API) | $3.00 | $15.00 | $0.30 |
| Self-hosted (open weights) | ~1.56 TB download; needs serious multi-GPU serving: most teams will rent hosted inference instead | ||
In the Kimi consumer app, the Highspeed variant is bundled into the Pro plan at $39/month. Migration is a one-line change: switch the model ID to kimi-k3.1-pro and keep everything else.
API Access (Kimi Open Platform): platform.kimi.ai
Benchmarks: What Holds Up Under Scrutiny
The strongest verified claim is coding-agentic performance. Moonshot reports 65.7% on Terminal-Bench 4.0; Artificial Analysis's independent run measured about 67%, the highest score on record on that test, ahead of Claude Opus 5.5. On the broader Artificial Analysis Intelligence Index, K3.1 Pro scores 55.1: fourth overall behind Claude Opus 5.5 (57.6), Claude Sonnet 5.5 (56.0) and Muse Wonder Full (55.8), and the highest of any open-weights model. That shrinks the best-open-vs-best-proprietary gap from roughly 12 points (K3 era) to 3.
Be skeptical of one number: Moonshot says its internal hallucination stress test improved from 23.4% to 8.3%, but Artificial Analysis measured a smaller drop, from 53% to 44%. Treat the 8.3% as a best-case marketing figure and keep fact-checking anything that matters.
Final Verdict: Should You Switch or Self-Host?
Switch your API workload to kimi-k3.1-pro if: you already run Kimi K3, or you run cost-sensitive coding agents where K3.1 Pro now tops Terminal-Bench at $2/$10: that is frontier-adjacent quality at a mid-tier price. Take Highspeed ($4/$20) only when end-to-end latency, not cost, is the binding constraint.
Self-host only if: you need air-gapped deployment, custom fine-tunes, or negotiated data terms, and you already operate multi-GPU inference infrastructure. The 1.56 TB download and MoE serving complexity mean most teams get strictly better economics from the API or a hosted provider. Stay on Claude or GPT if: you need the absolute frontier: Opus 5.5 still leads the overall index by a real margin, or your compliance policy rules out a Chinese-origin model.